Gaurav Pundir, Aseem Paranjape, Tirthankar Roy Choudhury

Building fast and accurate ways to model the distribution of neutral hydrogen during the Epoch of Reionization (EoR) is essential for interpreting upcoming 21 cm observations. A key component of semi-numerical models of reionization is the collapse fraction field $f_{\text{coll}}(\mathbf{x})$, which represents the fraction of mass within dark matter halos at each location. Using high-dynamic range N-body simulations to obtain this is computationally prohibitive and semi-analytical approaches, while being fast, end up compromising on accuracy. In this work, we bridge the gap by developing a machine learning model that can generate $f_{\text{coll}}$ maps by sampling from the full distribution of $f_{\text{coll}}$ conditioned on the dark matter density contrast $\delta$. The conditional distribution functions and the input density field to the model are taken from low-dynamic range N-body simulations that are more efficient to run. We evaluate the performance of our ML model by comparing its predictions to a high-dynamic range N-body simulation. Using these $f_{\text{coll}}$ maps, we compute the HI and HII maps through a semi-numerical code for reionization. We are able to recover the large-scale HI density field power spectra $(k \lesssim 1\ h\,{\rm Mpc}^{-1})$ at the $\lesssim10\%$ level, while the HII density field is reproduced with errors well below 10% across all scales. Compared to existing semi-analytical prescriptions, our approach offers significantly improved accuracy in generating the collapse fraction field, providing a robust and efficient alternative for modeling reionization.